Why everyone say Zen 5 is bad ?

[AusWolf]

Most gamers couldn't tell the difference between a 7600 non x and a 7950x3d if you threw a towel over the rig at 1440P -- but they will still overpay for the latter because cores and the FOMO of having the best of both worlds.

Look at these prices - the demand is high for this kind of thing.
View attachment 359881
Intel Prevails In Client CPU Sales, But Threadripper Pro Outsold Xeon Nearly 20:1 : Report | Tom's Hardware (tomshardware.com)

you push 32 cores at 3.6-4ghz zen5c into the consumer market, that chip will sell out even without vcache.

That's either not true, or whoever you mean by "most games" are incredibly stupid. Core count alone is not a thing for gaming. HWUB did a test not long ago showing that modern 6-cores run miles around old 8-cores in gaming. Sure, more cores are nice, but if the they're slow, then I'd rather have less of them, but faster.

 

[ARF]

HWUB did a test not long ago showing that modern 6-cores run miles around old 8-cores in gaming. Sure, more cores are nice, but if the they're slow, then I'd rather have less of them, but faster.

That's wrong. Six cores only would not be enough for the game and Windows and all of its main and background processes to run without micro-stuttering, and drops of the 0.1% lows.

 

[Melvis]

Won't bother to answer, your replies are beside the points made, reading comprehension is not your best asset.

Same! When ya that narrow minded and backed up by no proof there is no point!

 

[Mr_Engineer]

Oh for sure I know it has 0 chance of actually happening it's more of what I would like to see really from either cpu makers a low latency 12 core with lots of cache by the time we have it I will be saying 16 cores lmao because it will probably be 5 years from now at the soonest.

I believe AMD releasing a 12 core or even 16 core CCX is very possible/probable by the time Zen 6 is released (on 3nm or 2nm).

The thing is, in a few years, does the average home or office user benefit from that many cores, the answer is probably no. But for workstations and servers, a higher core CCX would be of great benefit for improved performance and efficiency.

 

[R0H1T]

A few more data points to look at ~

Spoiler

You should be able to get probably a lot more performance with some finetuning using PBO, undervolt, faster memory et al as compared to stock!

As always lowering TDP doesn't dent its performance way too much.

 

[phanbuey]

That's either not true, or whoever you mean by "most games" are incredibly stupid. Core count alone is not a thing for gaming. HWUB did a test not long ago showing that modern 6-cores run miles around old 8-cores in gaming. Sure, more cores are nice, but if the they're slow, then I'd rather have less of them, but faster.

It is true -- the question isn't "would you rather have more cores that are slower or less cores that are faster".... the question is - "would you pay for WAY more CPU than you actually need, or would notice" - the amount of people that answer yes to that with their wallet (myself included) massively eclipses the amount of people that balance their system to their actual needs.

Even today a common complaint among gamers with money is "It's really too bad the 7950x3d doesn't have vcache on both dies or I would have bought that" -- to @igormp points - that would be a product that is slower than the 7950x in production and slower than the 7800x3d in gaming, but people would still buy it like crazy...

Ryzen 5 7600 or 7 7800 x3d for 4K gaming | 4K DLSS and FSR | 4080 Super | 7900 XTX | 1440p (youtube.com)

I Upgraded To The 7800X3D And Regret It (youtube.com)

 

[dgianstefani]

Honestly the only attraction of Zen 5 for me is that X3D CPUs can be fully tuned now. Fluctuating clocks for no reason (they're under 60 C at all times) is something that irritates me.

I want to lock a 9950X3D at 5.2 GHz for both CCDs, turn off SMT, and use Process Lasso to put games only on cache CCD, and everything else on the other.

Will use same AM5 high performance heatspreader and delid setup I have now.

Otherwise Arrow Lake, which I'd prefer, but only if my friend wants my current rig.

If Zen 5 integrated the rumoured new packaging/IO die that Zen 6 comes with, it would be a no brainer, but Arrow Lake looks so much better on paper.

Could have avoided the poor reviews by simply moving every SKU up one notch, 9600X-8 cores, 9800X3D-8 cores VCache, 9900X-16 cores, 9950X3D-16 cores VCache. Noone could complain. But AMD...

Keep the dies that can't do 8 cores for server or OEM, where they're used anyway.

Noone could complain, simplified lineup. Most criticisms of Zen fixed, i.e. no gimped 6 core for gaming, no duplicate SKUs (normal and X3D), since there's no real compromise going with X3D now, because it can be OC'd, and should likely go to similar clocks due to the regression. Make a 9500 if they still want a six core, price at under $200, would fix the "no cheap AM5 chips". I mean, AMD released the 5900XT which is just a lower clocked 5950X anyway, so it's not like there's no precedent for 16 core xx90. Plus, would look good in core count/MT against Intel. I imagine they're eventually going to do one Zen CCD and one Zen "C" CCD, so 8+16 anyway, good interlude to that

Oh, and the compromise chip x900 wouldn't exist, since at the moment it has no benefit to consumer except for massively discounted since noone wants it. 8/16 are both better options. Yeah, this solution fixes most of my criticisms of Zen lineup, barring the IO die.

 

[igormp]

I feel like we're being attacked

x.com

x.com

 

[dgianstefani]

I feel like we're being attacked

x.com

x.com

"more cores from both Intel and AMD"

Difference is the Intel dual channel system, even on an architecture from two years ago, can get almost double the memory bandwidth, 125GB/s compared to what, 70?

 

[R0H1T]

And why does that matter in particular if results are (somewhat) decent? In fact AMD doing more (work) with less (bandwidth) is great.

 

[Assimilator]

And why does that matter in particular if results are (somewhat) decent?

Because attempting to justify AMD's shit IMC with "it's good enough bro" is just cringe.

 

[R0H1T]

He said Intel does 120 GBps which is a useless number without context, it's just synthertic reads(?) & means nothing without any workload attached to it. And as you can see even with ~50% more raw bandwidth Intel just sucks!

 

[chrcoluk]

There was a guy on reddit running a old 8 core skylake era chip, and he was 100% adamant that a modern 6 core wouldnt out perform it in gaming, I tried my best but moved on from trying to help him. His solution was just more and more cores, performance per core was meaningless.

 

[Paramount]

Zen 4 ended up looking better because people confuse ST with IPC. The ST performance uplift of Zen4 was quite good even better than Zen 3 that was already pretty good vs 3000.

I just didn't like how they were configured out of the box and the pricing of the platform for Zen4/am5..... I ended up getting a 7950X3D but it has it's quirks that while not hard to overcome are still annoying.

I have another AM5 board in a box I will likely just wait till the 9800X3D is cheap enough to want to use it Assuming Zen 6 skips AM5.

totally agreed
Zen5 is not a great upgrade if some one have zen4 cpu regardless of no. of cores ( with respect of the full support of AVX512 )
I also ended up and getting 7950x3d with great deal
7950x3d is impressive CPU for Gaming/multitasks with high efficiency

 

[igormp]

"more cores from both Intel and AMD"

Difference is the Intel dual channel system, even on an architecture from two years ago, can get almost double the memory bandwidth, 125GB/s compared to what, 70?

I guess you're comparing to those 7000+ sticks, right?
For DDR5-6000 the max theoretical bandwidth is... 6000*2*8=96GB/s. AFAIK Intel does manage to get reads pretty close to this limit, while AMD is stuck at like 75~80GB/s, similar for writes (intel gets similar write speeds to AMD at those frequencies IIRC).

But your point still stands since Intel's bandwidth does scale with higher frequency sticks, while AMD you'll be getting at most another 5GB/s or so.

And why does that matter in particular if results are (somewhat) decent? In fact AMD doing more (work) with less (bandwidth) is great.

Now imagine how better it could be if it weren't bottlenecked by memory.

But tbh, for the most users in this forum I guess games are what matters, and the IOD is still a bottleneck due to the latency penalty it incurs, and Zen 5 suffers a lot from that as well (hence why the V-cache might make more impact for it).

 

[dgianstefani]

I guess you're comparing to those 7000+ sticks, right?
For DDR5-6000 the max theoretical bandwidth is... 6000*2*8=96GB/s. AFAIK Intel does manage to get reads pretty close to this limit, while AMD is stuck at like 75~80GB/s, similar for writes (intel gets similar write speeds to AMD at those frequencies IIRC).

But your point still stands since Intel's bandwidth does scale with higher frequency sticks, while AMD you'll be getting at most another 5GB/s or so.

Now imagine how better it could be if it weren't bottlenecked by memory.

But tbh, for the most users in this forum I guess games are what matters, and the IOD is still a bottleneck due to the latency penalty it incurs, and Zen 5 suffers a lot from that as well (hence why the V-cache might make more impact for it).

Even at identical memory MT/timings Intel gets better numbers. Writes are also higher than AMD regardless of speed.

E.g. 7600 MT (doable on any Raptor Lake K series) gets 115-120 read and write, and 110 in copy, with around 50 ns latency.

 

[Wolverine2349]

Honestly the only attraction of Zen 5 for me is that X3D CPUs can be fully tuned now. Fluctuating clocks for no reason (they're under 60 C at all times) is something that irritates me.

I want to lock a 9950X3D at 5.2 GHz for both CCDs, turn off SMT, and use Process Lasso to put games only on cache CCD, and everything else on the other.

Will use same AM5 high performance heatspreader and delid setup I have now.

Otherwise Arrow Lake, which I'd prefer, but only if my friend wants my current rig.

If Zen 5 integrated the rumoured new packaging/IO die that Zen 6 comes with, it would be a no brainer, but Arrow Lake looks so much better on paper.

Could have avoided the poor reviews by simply moving every SKU up one notch, 9600X-8 cores, 9800X3D-8 cores VCache, 9900X-16 cores, 9950X3D-16 cores VCache. Noone could complain. But AMD...

Keep the dies that can't do 8 cores for server or OEM, where they're used anyway.

Noone could complain, simplified lineup. Most criticisms of Zen fixed, i.e. no gimped 6 core for gaming, no duplicate SKUs (normal and X3D), since there's no real compromise going with X3D now, because it can be OC'd, and should likely go to similar clocks due to the regression. Make a 9500 if they still want a six core, price at under $200, would fix the "no cheap AM5 chips". I mean, AMD released the 5900XT which is just a lower clocked 5950X anyway, so it's not like there's no precedent for 16 core xx90. Plus, would look good in core count/MT against Intel. I imagine they're eventually going to do one Zen CCD and one Zen "C" CCD, so 8+16 anyway, good interlude to that

Oh, and the compromise chip x900 wouldn't exist, since at the moment it has no benefit to consumer except for massively discounted since noone wants it. 8/16 are both better options. Yeah, this solution fixes most of my criticisms of Zen lineup, barring the IO die.

Good thoughts and good points.

You say you are interested in Zen 5 3D because it will be more tunable. Well hopefully. I remember the same was said about Zen 4 X3D and it was barely more tunable and still locked when it finally came out.

Hopefully it really is different with Zen 5 X3D. Though I hear the extra cache is so sensitive to extra heat that it can fry so easily so AMD may be forced to restrict this as well? Is there reason to believe it will truly be different this time with Zen 5 3D?

And you state you would prefer Arrow Lake as it looks better but only if friend wants current rig? Meaning do not want to have to sell it and get new mobo and extra cost if you can just pop in Zen 5 X3D CPU when it comes out?

 

[dgianstefani]

Good thoughts and good points.

You say you are interested in Zen 5 3D because it will be more tunable. Well hopefully. I remember the same was said about Zen 4 X3D and it was barely more tunable and still locked when it finally came out.

Hopefully it really is different with Zen 5 X3D. Though I hear the extra cache is so sensitive to extra heat that it can fry so easily so AMD may be forced to restrict this as well? Is there reason to believe it will truly be different this time with Zen 5 3D?

And you state you would prefer Arrow Lake as it looks better but only if friend wants current rig? Meaning do not want to have to sell it and get new mobo and extra cost if you can just pop in Zen 5 X3D CPU when it comes out?

1.3 VCore is the only limitation.

Since you can't really go higher than that anyway without issues even on normal Zen 5, it won't be the problem.

Zen 4 X3D wasn't really "more tunable", you could just use PBO.

 

[oxrufiioxo]

Good thoughts and good points.

You say you are interested in Zen 5 3D because it will be more tunable. Well hopefully. I remember the same was said about Zen 4 X3D and it was barely more tunable and still locked when it finally came out.

Hopefully it really is different with Zen 5 X3D. Though I hear the extra cache is so sensitive to extra heat that it can fry so easily so AMD may be forced to restrict this as well? Is there reason to believe it will truly be different this time with Zen 5 3D?

And you state you would prefer Arrow Lake as it looks better but only if friend wants current rig? Meaning do not want to have to sell it and get new mobo and extra cost if you can just pop in Zen 5 X3D CPU when it comes out?

As much as I like my 7950X3D having to manually tell it what to do for every program is annoying, them just fixing the scheduling issues so I don't have to do it manually on the 9950X3D would be a win in my book.

Sure you can fix it with process lasso but it just being plug and play would be so much better.

Given that the 9950X needs to almost be treated like an X3D part with scheduling and has over double the ccd to ccd latency I'm not holding my breath.

Even if you could OC it I wouldn't so that doesn't interest me I also doubt it'll make much difference other than to consume way more power for an insignificant amount of extra MT performance. Some people are just too OCD about clocks etc lol. Good for them.

Looks like there was a leak for the core ultra 9 cpu looks about 10% faster in ST and MT vs the 14900k which if comes at 25-30% less power and scales to other workloads not just geekbench not bad but still kinda meh. Obviously need to wait for real reviewes either way.

 

[Wolverine2349]

As much as I like my 7950X3D having to manually tell it what to do for every program is annoying, them just fixing the scheduling issues so I don't have to do it manually on the 9950X3D would be a win in my book.

Sure you can fix it with process lasso but it just being plug and play would be so much better.

Given that the 9950X needs to almost be treated like an X3D part with scheduling and has over double the ccd to ccd latency I'm not holding my breath.

Even if you could OC it I wouldn't so that doesn't interest me I also doubt it'll make much difference other than to consume way more power for an insignificant amount of extra MT performance. Some people are just too OCD about clocks etc lol. Good for them.

Looks like there was a leak for the core ultra 9 cpu looks about 10% faster in ST and MT vs the 14900k which if comes at 25-30% less power and scales to other workloads not just geekbench not bad but still kinda meh. Obviously need to wait for real reviewes either way.

Yes bingo good points. Which is why AMD parts suck with dual CCDs. Plus the severe cross latency penalty.

It sucks there is no good more than 8 core option.

Intel has thread director, but still Big.Little and sometimes scheduling quirks with that but not as bad as AMD but still not ideal and what I would like. And that really bad because AMD 9950X and 9900X are like dual socket 9700X and 9600-X systems and since when did dual socket systems need special scheduling despite the severe latency hit. The 5000 series did not even though latency hit sucked crossing CDs cause it did not matter for most productivity separate NUMA nodes unlike gaming. The 9000 and even early on 7000 series was another animal. But all dual CCD parts and scheduling and latency sucks especially latency in regards to 5000 though scheduling seemed ok.

 

[oxrufiioxo]

Yes bingo good points. Which is why AMD parts suck with dual CCDs. Plus the severe cross latency penalty.

It sucks there is no good more than 8 core option.

Intel has thread director, but still Big.Little and sometimes scheduling quirks with that but not as bad as AMD but still not ideal and what I would like. And that really bad because AMD 9950X and 9900X are like dual socket 9700X and 9600-X systems and since when did dual socket systems need special scheduling despite the severe latency hit. The 5000 series did not even though latency hit sucked crossing CDs cause it did not matter for most productivity separate NUMA nodes unlike gaming. The 9000 and even early on 7000 series was another animal. But all dual CCD parts and scheduling and latency sucks especially latency in regards to 5000 though scheduling seemed ok.

It's around 70ns on 7000 which while not good isn't terrible on ryzen 9000 it's 170ns smh

Hopefully a bug not sure how it could have regressed that bad.

 

[Wolverine2349]

It's around 70ns on 7000 which while not good isn't terrible on ryzen 9000 it's 170ns smh

Hopefully a bug not sure how it could have regressed that bad.

70ns is pretty bad considering core to core latency is like in the teens. Its not consistent. That is the problem. Its a massive jump.

Intel 10900K had 10 cores all on same ring and core to core latency consistent across board. Though I have heard that even 10900K had instability issues cause of too many cores on the ring and WHEAs. How true is that regarding 10900K. Some say its not possible without errors to go much beyond 8 P cores on a single ring? Is that true or no? Or is up to 12 fine if done correctly?

 

[igormp]

Honestly the only attraction of Zen 5 for me is that X3D CPUs can be fully tuned now. Fluctuating clocks for no reason (they're under 60 C at all times) is something that irritates me.

I want to lock a 9950X3D at 5.2 GHz for both CCDs, turn off SMT, and use Process Lasso to put games only on cache CCD, and everything else on the other.

What do you mean by that?
AFAIK, the 9000 series still has the same "issue" the 7000 series had, which is that both CCDs share the same voltage rail, but for the x3D chips the voltage ceiling is lower, which ends up capping the non V-cache CCD.

1.3 VCore is the only limitation.

Since you can't really go higher than that anyway without issues even on normal Zen 5, it won't be the problem.

Zen 4 X3D wasn't really "more tunable", you could just use PBO.

We need to see the voltage ceiling for the V-cache chips, but if nothing changes from the 7000 series I don't think you'll achieve what you want:

Raphael only has one VDDCR voltage rail powering the CPU cores in the CCDs via the VDDCR_CPU rail and the cache via the VDDCR_VDDM rail. As I discussed in my Raphael launch content, it means that when any core in CCD0 is active, it will also restrict the maximum voltage for the cores in CCD1. That will also constrain its operating frequency.

Note that the voltage limits are the absolute limits at zero load. When the CPU is under a substantial load, the Precision Boost algorithm will further restrict the allowed voltage. The maximum voltage in an all-core workload with active cores in CCD0 is a mere 1.0V. That’s similar to the maximum voltage for a stock non-X Ryzen 7000 processor. The maximum voltage in an all-core workload for CCD1 is 1.2V

The V/F curves for both CCDs are similar, meaning they run approximately the same frequency at the same voltage. In an all-core workload with cores in both CCDs active, the frequency is about 4.6 GHz at 1.0V. However, in that same workload, the cores in CCD1 can run at 5.2 GHz at 1.2V.

From: https://skatterbencher.com/2023/02/27/skatterbencher-56-amd-ryzen-9-7950x3d-overclocked-to-5900-mhz/

But since all you want is 5.2GHz, this may be doable.

 

[oxrufiioxo]

70ns is pretty bad considering core to core latency is like in the teens. Its not consistent. That is the problem. Its a massive jump.

Intel 10900K had 10 cores all on same ring and core to core latency consistent across board. Though I have heard that even 10900K had instability issues cause of too many cores on the ring and WHEAs. How true is that regarding 10900K. Some say its not possible without errors to go much beyond 8 P cores on a single ring? Is that true or no? Or is up to 12 fine if done correctly?

Why I keep saying improving the IF should be the number 1 priority for amd.


On Intel no idea why they haven't scaled past 8 P cores. They must have their reason from a technical standpoint.

I feel like 8 cores have turned into the new quad cores with being stuck on them lol. With only multiple ccd or e cores being the option if you want more.

 

[Wolverine2349]

Why I keep saying improving the IF should be the number 1 priority for amd.


On Intel no idea why they haven't scaled past 8 P cores. They must have their reason from a technical standpoint.

I feel like 8 cores have turned into the new quad cores with being stuck on them lol. With only multiple ccd or e cores being the option if you want more.

Yes and it absolutely sucks and is terribly sad and unacceptable. And this is with supposedly good competition between Intel and AMD in mobile and desktop markets. and even server markets despite AMD having upper hand in that

There is not even a good HEDT option even if your wiling to pay top dollar. The Xeon Workstation sucks for gaming and requires ECC RAM and has severely gimped IPC compared to client Golden Cove and that platform has enterprise features detrimental to gaming beside just costing more.

Back in the time where there was 0 competition between Intel and AMD and Intel had it all, it felt as if there were more and better options in an era where few if any games scaled beyond 4 cores.

I mean if you were willing to pay top dollar, Intel had HEDT and Haswell E and Broadwell E from 2014 to 2016 and you could not only get 8 cores on a ring, but oh 10 on a single ring if willing to pay top dollar for the CPU. And it rocked for gaming even if it was overkill in core count. Had same IPC and clocks as mainstream 4 core Haswell and no ECC RAM requirement.

Why no such option now. They do not seem to care about us DIYers anymore. DIYers and desktop market not as niche as people think as MicroCenter would not have expanded to Indianapolis and Charlotte and planned to Miami if that was the case.

Its a shame Intel forcing Big.Little with e-cores down our throats. Please intel release a 12 + 0 die for Arrow Lake in October. You have an immediate buyer in me. And you would put a bigger dent in AMD sales even more given Zen 5 flop.

Its actually worse now than then because then there were good 8 cores options in HEDT if you were willing to pay unlike the crappy Enterprise gimped and overloaded crap and bad IPC of SPR.

Plus back then like 0 games scaled beyond 4 cores but even if some rare ones did a great options in Intel HEDT.

Now there are no good option even in Workstation Threadripper or HEDTs and willing to pay top dollar and the consumer and mainstream platforms the later of which without scheduling gimmicks and quirks and latency penalty crap and this is in an era where games can scale beyond 8 cores more so now in 2024 than games could scale beyond 4 back in 2014 to 2016.